AI developer tools · What shipped

Microsoft released BenchPress, a machine learning framework for predicting AI model performance across benchmarks without running full evaluations. The tool compiles a public score matrix of 84 frontier models on 133 benchmarks and uses logit-space bias-decomposed rank-2 matrix completion to predict missing scores with a median absolute error of 4.6 points. The framework achieves 92.1% preservation of pairwise model rankings using only five probe benchmarks, reaching 5.0 error on brand-new releases. BenchPress is available as open-source code and dataset (MIT and CDLA-Permissive-2.0 licensed) on GitHub with a living dataset maintained on Hugging Face, supporting 189 frontier LLMs from 28 providers and 316 benchmarks across 59 categories.

Alibaba released Qwen3.8-Flash, an open-weight multimodal Mixture-of-Experts model featuring a 125B-parameter main model with 6B parameters activated per token. The model performs competitively against DeepSeek-V4-Flash and Claude-Opus-4.6 across multiple benchmarks including SWE-bench Pro, CoWorkBench, Toolathlon Verified, MathVision, AndroidWorld, and ERQA. Qwen3.8-Flash supports 262K context tokens natively (extendable to 1M), with weights available on Hugging Face and ModelScope, and API access priced at 0.16 USD per million input tokens and 0.47 USD per million output tokens. The model introduces architectural innovations including Gated DeltaNet attention combined with Qwen Sparse Attention, Gated Residual mechanisms, N-gram embeddings, and the Muon Optimizer to improve computational efficiency. It requires only one-ninth the training resources of Qwen3.7-Plus while delivering superior performance in coding and office tasks, serving as a preview for the upcoming Qwen4 series.

Unsloth released quantized versions of Qwen3.8 models with new developer tools and optimizations. Qwen3.8-27B runs on 16-19GB VRAM with vision and reasoning capabilities across 256K context window, available as GGUFs via Unsloth with Dynamic V3.0 achieving 10% more accuracy at the same size. Qwen3.8-2.4T-A95B (95B active parameters) was quantized to 397GB using new 1-bit data-types (UD-IQ1_XXXS at 1.1875 bits per weight), reducing storage 91% from full precision. Unsloth released NVFP4 quants running 1.5x faster than BF16 with 92-97% accuracy retention on Blackwell GPUs, available through vLLM, llama.cpp, and Unsloth Desktop with tool calling support for agentic coding tasks.